Retell AI
AI · AI
Voice and chat agents, phone numbers, and calls with their transcripts.
Connecting Retell AI in practice
Retell’s real ceiling is concurrent calls per workspace — 20 by default on pay-as-you-go — and an inbound call that finds no free slot waits about 40 seconds before being sent to a configured fallback_number or ending as concurrency_limit_reached. Switching on concurrency burst lets calls over the limit through, but adds $0.10 a minute to the entire duration of every burst call.
Scopes that draw scrutiny
Monitor → History (Read or Edit)Retell’s own permission group covering call and chat history, transcripts, recordings and call metadata — the recorded conversations themselves.Deploy → Call (No Access or Edit only)Creates and manages web, phone and batch calls plus live-call controls. It is action-only with no read level, so granting it at all means the key can place calls that cost money.Deploy → Phone (Read or Edit)Phone numbers, A2P campaigns, business profiles and branded-call verification, which is where telecom registration and fraud exposure live.Build → Agent (Read or Edit)Agents, conversation flows, Retell LLMs and knowledge bases — editing these changes what every future call says.Unrestricted API keyA key created without "Restrict permissions" has full access to every endpoint that supports API-key auth, which is the default.
Rate limits
- Concurrency is per workspace, not per account: 20 concurrent voice calls by default on pay-as-you-go, adjustable from the dashboard.
- Burst limit is the lower of 3× the concurrency limit or the limit plus 300; burst calls cost an extra $0.10/min for the whole call.
- CPS (calls per second) is set separately per telephony path — Telnyx, Twilio and custom telephony each have their own card and range.
- Inbound calls queue for roughly 40 seconds at the limit, then fall back or end with concurrency_limit_reached.
- Maximum call duration is 1 hour by default and can be raised to 2 hours in agent settings.
Who has to approve
Retell gates the limits themselves by role: adjusting concurrency and CPS is available to the Admin and Developer roles, while reserving inbound capacity and toggling concurrency burst change workspace settings and require Admin. Only one API key at a time can be designated the webhook key.
Worth knowing
- Inbound, outbound, web and batch calls all draw on the same concurrency pool, so a batch campaign starves the support line unless reserved_inbound_concurrency is set.
- Retell LLM prompts cap at 32,768 tokens and anything over 4,000 tokens is charged extra, so a long system prompt is a per-call cost rather than a one-off.
- Webhook verification uses the X-Retell-Signature header against the raw request body, and Retell also documents IP allowlisting — a re-serialized body fails verification.
- Data retention is configured per agent, so transcripts and recordings for one agent can outlive another’s by design.
Checked against Understand concurrency and limits (2026-09-29), Manage API keys and permission scopes (2026-09-29), Secure the webhook (2026-09-29), Data retention policy (2026-09-29)
What an agent can do
Each action is granted on its own. An agent allowed to read is not thereby allowed to write, and the scope beside each row is what the acting user must have connected for it to run at all.
Often connected alongside
Put Retell AI behind one governed endpoint.
Same permissions, same audit trail, whatever else you connect next.